Classification of Sparse and Irregularly Sampled Time Series with Convolutional Neural Networks

Analysis of sparse and irregularly sampled time series is an important task with prominent applications in various domains such as medicine, manufacturing, social networks and environmental sciences. In this paper, we focus on convolutional neural networks
for the classification of the aforementioned type of time series. We point out that dynamic time warping (DTW) is suitable in case of such time series and present a new convolutional block, dynamic time warping convolution with sparse input (DConv). We propose to use DConv for the classification of sparse and irregularly sampled time series.We also consider networks with sparsity-invariant convolution, an operation which is known in the domain of 3D reconstruction, but new in the domain of time series classification, and use neural networks with sparsity-invariant convolution as strong baselines. We perform experiments with fully-convolutional neural networks (FCNs) and residual networks (ResNet) on publicly available time series datasets. The results show that neural architectures with the proposed convolutional block outperform their counterparts, in most cases statistically significantly. We made our implementation publicly available at https://github.com/kr7/dconv.